DALL·E 2

What it is

DALL·E 2 is an image generator that OpenAI announced on 6 April 2022 ^[1]. It is described in the paper "Hierarchical Text-Conditional Image Generation with CLIP Latents", by Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu and Mark Chen ^[2]. It is the second text-to-image system from OpenAI, following the original DALL·E.

Origins

The system's approach has two main parts. A prior turns a caption into a CLIP image embedding, which is a numerical representation of an image's meaning. A diffusion decoder then generates the image from that embedding. This design enables variations of an existing image and language-guided editing ^[2].

Diffusion models start from random noise and gradually refine it into an image, guided by the input. Conditioning on CLIP embeddings means the image is steered by a representation that links text and pictures.

The model uses 3.5 billion parameters, fewer than the original DALL·E, which had 12 billion ^[1].

Main features

The model supports image variations, which produce new versions of a given image, and language-guided edits, which change parts of an image according to a text instruction ^[2]. An outpainting feature, which extends an image beyond its borders, was introduced on 31 August 2022 ^[1].

DALL·E 2 entered beta on 20 July 2022. The beta gave invitations to 1 million people from the waitlist, with free monthly credits. Additional credits were sold in increments of 115 generations for $15 ^[1][3].

The waitlist was removed on 28 September 2022, so anyone could sign up after that date ^[1].

API and pricing

API access followed in early November 2022, with images priced per image according to resolution ^[1].

Retirement

OpenAI announced its API deprecation on 14 November 2025. The dall-e-2 model was shut down on 12 May 2026 ^[4]. Deprecation means the model is scheduled to be retired, and the shutdown means it can no longer be used.

Place in the lineage

DALL·E 2 was the successor of the original DALL·E, and it was succeeded by DALL·E 3 in 2023 ^[1]. Its main design choice is the combination of CLIP embeddings and a diffusion decoder.

For readers, DALL·E 2 is a useful milestone. It moved text-to-image generation from a research paper and a waitlist to a paid service with a public API.